Machine Learning Automation: A Game-Changer for the Insurance Industry

Transcription

1 Machine Learning Automation: A Game-Changer for the Insurance Industry Satadru Sengupta, General Manager, DataRobot Insurance Industry

2 Satadru is the General Manager of Insurance Practice at DataRobot where he works with different insurance companies (Life, Health, and P&C/ General) to integrate the DataRobot Machine Learning Automation platform in their problem-solving environment across different functions (e.g. UW, pricing, audit, claims, operations). Previously, Satadru worked with AIG Science Team as a Senior Manager and led quantitative modeling for the Global Distribution Team. Prior to that, he worked with Liberty Mutual Insurance and Deloitte Consulting. Satadru is a frequent speaker at different insurance conferences (including SOA and CAS).

3 Introduction Autonomous cars. Robots that understand human speech. Computers that can play games. Stories about machine learning in the press feature exotic applications, remarkable things that could be useful in a few years. But that isn t the full story. Right now, insurers use machine learning for very practical things: Managing broker business Optimizing direct marketing Understanding quote conversion Computing optimal pricing Detecting fraud Claims triage Predicting litigation Targeting inspections and audits Forecasting claims Retaining customers Recalibrating prices Insurers use machine learning to make business processes smarter. And, they measure the gains from doing so in the millions of dollars added to the top and bottom lines. The insurance industry invented data-driven analysis, and most insurance executives understand statistical modeling. Machine learning is a step beyond statistical modeling. It leverages today s cheap and pervasive computing power, massive databases, and modern software architecture, all directed towards one objective: better predictions without sacrificing interpretability. Like any innovation, there are some myths and misunderstandings about machine learning. You don't want to get stuck in these and miss out on the benefits of this technology. We'll take a critical look at four common objections you may hear in your organization. Finally, we ll introduce you to DataRobot, a machine learning and data science company headquartered in Boston. DataRobot helps insurers take machine learning out of the research lab and put it to work. We'll also introduce you to some DataRobot customers, who use machine learning to build better mortality tables, predict claims costs, reduce customer churn, support market expansion and calculate equitable premiums. To learn more, keep reading.

4 The Power of Machine Learning Machine learning is in the news. Google says it is rethinking everything around machine learning. In Slate, tech writer David Auerbach argues that machine learning reshapes how we live. In Harvard Business Review, Mike Yeomans writes that every manager should know about machine learning. If you are an insurance executive, these claims may seem like hype. Insurers invented some of the most advanced statistical techniques more than a hundred years ago. But consider how these insurers use machine learning today. Progressive Insurance uses machine learning to predict claims from telematics and geospatial data. Zurich Insurance uses machine learning to support marketing, fraud detection, and claims management. Transamerica uses machine learning to recommend products to customers. One of the largest U.S. insurers uses machine learning to deliver value across lines of business and functions. The company has many applications that rely on machine learning. In the worker s compensation business, managers use predictions of claims severity to target investigations and interventions. Using individual-level predictions, sales managers channel business to brokers and agents with the highest retention and submission efficiency. Claims fraud detection models developed in-house detect twice as much fraud as leading vendor systems. Let's look at the different ways insurers use machine learning. Submission Prioritization. Insurers use machine learning to predict premiums, conversion, and losses for the policies that brokers submit based on the data available on the first day. This practice helps underwriters focus on the most valuable business. Detecting good risks early in the process enables insurers to make better use of underwriters time and delivers a huge competitive advantage. Direct Marketing. Insurance marketers use machine learning in many ways. Machine learning can identify the best prospects for a product or offer. It can also determine the best offers for individual customers. Insurers use machine learning to predict the profitability of new products and predict market size. They use it to profile customers, model propensity to buy products, and measure price elasticity.

5 For the insurance sector, we see machine learning as a fundamental game-changer since most insurance companies today focus on three main objectives: improving compliance, improving cost structures and improving competitiveness. Machine learning can form at least part of the answer to all three. Gary Reader, KPMG s Global Head of Insurance Conversion. Insurers monitor the conversion rate on customer quotes. A low conversion rate may signal that the price is too high. Machine learning helps insurers understand conversion drivers. This knowledge helps the insurer target price adjustments. Robust machine learning algorithms also yield more accurate pricing with fewer factors. This capability makes quoting faster and improves conversion. Optimal Pricing. Many insurers quote rates with simple models that regulators understand. Behind the scenes, insurers use machine learning to build optimal pricing models. These models help the company make better pricing decisions. Targeting Inspections and Audits. Insurers use machine learning to identify policies with a high risk of misreporting and use the results to target inspections and audits. This procedure helps avoid mispriced risks without disturbing valued customers. Fraud Detection. Insurance fraud is a growing problem. Insurers use machine learning to shift through claims and identify those that warrant deeper investigation. Claims Triage. Insurers can save millions of dollars in claim costs through proactive management: fast settlement, targeted investigations, and case management. Machine learning helps insurers detect complex claims early in the lifecycle, so they can triage cases accordingly. Predicting Litigation. Legal defense costs drive total claims cost for most commercial insurance products, including workers compensation, product liability, and medical liability. At First Notice Of Loss (FNOL), insurers use machine learning to auto-adjudicate claims and triage them accordingly. Doing so early in the process significantly reduces claims cost. Claims Forecasting. Calculating reserves is a balancing act between prudence and profitability. Insurance executives need accurate loss predictions so that they can set reserves appropriately. Machine learning delivers the high-quality predictions insurers need for smart decisions. Customer Retention. Customers are valuable. With machine learning, insurers identify causes of customer attrition. This knowledge helps the insurer take preventive action. Machine learning also helps insurers target at risk customers. Let s summarize: Machine learning delivers more accurate predictions than traditional analysis or human judgment. Modern techniques make these predictions easy to understand and transparent. With better predictions, managers make smarter decisions. Smarter decisions produce more revenue, lower costs, and a better bottom line. What can machine learning do for your company?

6 From Statistics to Machine Learning Insurers rely heavily on statistical models. In regulated markets, insurers use methods like Generalized Linear Model (GLM) to compute individual price quotes. Regulators understand GLM and similar techniques and tend to trust the results. Generalized Linear Model (GLM) is an advanced form of linear regression modeling. In ordinary linear regression, the measure we want to predict (called the response variable) should conform to a bell curve, or normal distribution. Many real-world problems do not satisfy this condition. GLM supports many other statistical distributions. Actuaries and analysts use tools like GLM to measure the relationship between the characteristics of an insurance policy and the expected losses from that policy. (For example, an auto insurer would gauge the effect of features of the insured's driving record, type of vehicle, garaging location and many other things.) The analyst trains the model with historical data. The insurer then uses that model to generate loss predictions for prospective new policies and prices the business accordingly. Statistical methods fit a limited number of mathematical models to the data; the analyst chooses a model and checks to see how well it fits the data. For simple problems and small data sets, statistical methods work reasonably well. They are also relatively easy to solve, which was an important consideration before the advent of cheap and pervasive computing. In a seminal paper published in 2001, Leo Breiman, Professor of Statistics at the University of California, Berkeley, explained differences between statistics and machine learning. Machine learning algorithms do not start with a particular mathematical model. Instead, they use a variety of methods to search for the best way to solve the prediction problem. Consequently, they deliver more accurate predictions than statistical methods because they can detect very complex patterns in the data. (Practitioners call thes intricate patterns nonlinear relationships and interaction effects.) As analysts seek to enrich predictive models with new data, datasets are getting wider there are more variables in the data that may have a relationship with the thing we want to predict. Ten years ago, insurance analysts rarely worked with more than a hundred variables; today, many analysts routinely sift through thousands. Machine learning techniques work very well with these wide datasets. In many cases, insurers have too little data, not too much data. For applications like submission prioritization, price quotes, and litigation prediction, minimal data is available. Machine learning extracts maximum predictive power from thin data. The discipline of machine learning is not new; many algorithms date back to the 1940s. However, few organizations used machine learning until recently, because the techniques need a lot of computing power to deliver results that are markedly better than statistical methods. Several converging trends contribute to the current surge of interest in machine learning. The most important of these is Moore s Law: the radically declining cost of computing. Today, every business has more computing power available than anyone dreamed about as little as ten years ago, at a fraction of the cost.

7 Suppose that you are driving to another city and you have exactly ten gallons of fuel. Should you take the gas guzzler or the hybrid? Most people would choose the fuel-efficient car to so they can travel the greatest distance on the available fuel. The same holds for predictive analytics: if you don t have much data, select the most efficient algorithm you can find. That s the power of machine learning. Cheap and pervasive computing power drives an explosion of innovation in machine learning. Scientists, researchers, and practitioners introduce new algorithms at a dizzying pace. Each new algorithm is a little better than its predecessors, but the cadence of innovation is so rapid that the cumulative impact is radical. In some fields, we see tenfold improvements in speed and scalability every six to twelve months. With cheap computing and constant innovation, machine learning practitioners accumulate knowledge and experience. They are better able to meet business requirements because they know how to configure and tune algorithms to produce the best possible predictions. Statistical methods are sound techniques with a solid track record. Machine learning methods deliver more accurate predictions than statistical methods, and they work better with "big and wide" data. Not that long ago, only a few researchers with supercomputers could benefit from machine learning. Today, when your mobile phone has more computing power than a 1970s-era mainframe computer, machine learning is available to everyone. Four Myths About Machine Learning But wait! you may think. Machine learning is complicated. That s fine for big companies like AIG and Zurich, but there s no way we ll ever be able to afford that. Let s talk about that. Myth #1: Machine learning is opaque. Regulators prefer statistical models because they believe they can understand them simply by inspecting the code. Machine learning models, in contrast, seem very hard to interpret. This perception is due, in part, to simple familiarity it s easier for people to understand techniques they already know. It is also due to the greater complexity of machine learning models; machine learning models deliver more accurate predictions because they fit our complex world better than statistical models. While it is harder to understand a machine learning model by inspection, it is still possible to know how it behaves, and how it produces predictions. If we focus on how well a machine learning model predicts, and not what it looks like, we can understand how well it will perform when we use it in a production environment. Moreover, there are graphical tools available today that make it easy for a business user to see how a machine learning model behaves, and tools that deliver explanations for each prediction.

8 Myth #2: We will never be able to hire enough experts to use machine learning. Machine learning used to be the exclusive domain of experts called data scientists, who combine business domain knowledge, analytics knowledge, and the ability to write code. Good ones are hard to find, hire and retain; The Wall Street Journal, The Chicago Tribune, and many others all note the shortage. Data scientists are so rare that Harvard Business Review suggests that you stop looking or lower your standards. The good news is that modern machine learning tools are much easier to use today than just a few years ago. Your organization already has domain experts: actuaries, claims managers, customer service reps, marketing managers, and underwriters. With the right tools and training, anyone can contribute to machine learning projects. Myth #3: Machine learning projects take forever. Managers recently surveyed by Gartner said that time to value is one of their biggest problems with machine learning. Those managers reported that it takes an average of 52 business days for their team to build a predictive model. (Some said it took months.) When your business needs the benefits of machine learning, weeks or months can seem like forever. Fortunately, modern machine learning software is faster than was in the past. It works on distributed platforms so that you can harness more computing power. With deployable code and prediction APIs, getting models into production takes a lot less time. Myth #4: Machine learning is a headache for IT." IT executives who work with legacy tools for machine learning and predictive analytics know how expensive and troublesome they are. Legacy software for advanced analytics is costly to own and use. The software is expensive, often requires non-standard hardware, constant maintenance, and specially trained people. Fortunately, modern machine learning platforms reflect enterprise computing standards. They work with widely used data platforms, conform to security standards, and run efficiently on commodity hardware. The best examples leverage open source software for ease of integration and low cost of ownership. DataRobot for Machine Learning DataRobot, a global data science and machine learning company headquartered in Boston, Massachusetts, helps insurers and other industries around the world leverage the power of machine learning. DataRobot provides a platform for users of all skill levels to produce accurate predictions in a fraction of the time needed using conventional tools and methods. DataRobot automates the machine learning process, from data to deployment. DataRobot supports the advanced machine learning techniques used in the insurance industry today, including Generalized Additive Models (GAM), and algorithms that support offset, exposure and the Tweedie family of loss functions. To leverage the most innovative techniques, DataRobot uses open source tools like Apache Spark, H2O, Python, R, TensorFlow, Vowpal Wabbit and XGboost.

9 DataRobot for Machine Learning Empowers actuaries. DataRobot s machine learning automation platform perfectly complements actuaries' deep business acumen and mathematical background. It combines an intuitive web interface and graphical model assessment tools with flexible programming APIs, so users can work with the tools they know. DataRobot supports advanced tuning, and it supports custom extensions for the most sophisticated analysis. Enables collaboration. Actuaries, underwriters, claims adjusters, fraud investigators, and other stakeholders can work together on machine learning projects with DataRobot. Collaboration moves machine learning out of the skunk works and into the front office. With a clear definition of the business problem and direct feedback from stakeholders, machine learning projects deliver better results. Interpretable and transparent. With DataRobot, every step in the modeling process is visible and reproducible. DataRobot offers extensive facilities for understanding the behavior of the predictive models it builds. It also provides aids to the business user or stakeholder seeking to understand and interpret the model. DataRobot explains its predictions by highlighting the most influential factors, so your team can explain results to customers, executives and regulators. Solves many problems. Insurers use DataRobot for machine learning across many different business functions, such as Actuarial, Claims, Fraud, Marketing, and Underwriting. DataRobot is a general purpose machine learning platform you can use to transform your business across different products, functions, and regions. It helps you avoid the integration, training, and support headaches of using many different "point" solutions in various parts of your business. Reduces time to value. DataRobot is fast. Its modern software architecture uses distributed computing to run many experiments in parallel. It automates routine tasks like data discovery, feature engineering, and model assessment. Once your organization approves a model, DataRobot offers multiple options for deployment, including native scoring, exportable prediction code and prediction APIs. Enterprise-ready. DataRobot works with your data wherever it resides: relational databases, Hadoop, text files and many other sources. It runs on free-standing clusters, in Hadoop under YARN or as a managed service in the cloud. In Hadoop, DataRobot uses native security, data provenance and application management services. For low maintenance and easy integration, DataRobot uses the most current software development practices, including microservices and containerization. DataRobot delivers the best predictions possible with your data. It packages data science expertise, best practices, and world-class Insurance industry experience in a machine learning automation workflow. With built-in expertise and guardrails, you can be confident that your users will consistently produce valid predictive models.

10 Customer Success Stories Tier One Reinsurer A Tier One reinsurer with billions in gross written premiums uses DataRobot to underwrite impaired Life patients, for mortality prediction, and for segmentation. The actuarial team used DataRobot to build a breast cancer mortality model that is 25% more accurate than an existing model constructed with statistical methods. From this model alone, the company estimates millions of dollars per year in claim cost savings. Global Life and Health Insurer A global life and health reinsurance company with trillions of dollars of reinsurance in force uses DataRobot for actuarial and marketing use cases. The actuarial team uses DataRobot to predict mortality risks among cancer patients in remission. A robust mortality model simplifies underwriting since low-risk applicants can avoid manual medical verification. It also reduces claims costs by identifying high-risk patients. With DataRobot's built-in best practices, the company built a mortality model that was significantly more accurate than an existing model constructed with statistical methods. International Insurance Company A leading international insurance and financial services firm launched a program to transform its business. The company targeted four areas: claims costs, customer churn, market expansion, and calculating equitable premiums. With DataRobot, the insurer s transformation teams used machine learning to build accurate predictive models to support smart decisions across all four functions. In Claims, managers can predict the most severe claims, and target interventions accordingly. With better predictions of customer churn, the firm can intervene to prevent the best and most profitable customers from leaving. For market expansion, the insurer now uses machine learning to understand the profit potential in different market segments. With machine learning, the company has a more precise understanding of customer risk. Regional Workers Compensation Carrier A U.S.- based mutual insurance company with 300 million in net premiums is one of the largest writers of workers' compensation coverage in its region. The company competes with billion-dollar carriers and needs every possible advantage it can find. The company relied on outside consultants and vendors to deliver predictive models because it did not make economic sense to staff a large team of insurance analysts. With DataRobot, the insurer's business analysts could build predictive model without outside consultants. Interestingly, they discovered that the models they created with DataRobot outperformed the models delivered by outside experts. By "insourcing" insurance analytics, the company reduced spending on consulting fees and deployed better predictive models. For more details visit or contact us at info@datarobot.com